wsr

Next inspection iteration: literal runner-label lists

Status: delivered. PR #25 merged as 91fd3d5 after all four required checks passed on 722b164, with no review findings. The merged CLI fixture was verified; local main was clean. Baseline: PR #24 merged as b22b339 on main; explicit step identities are delivered. The owner instructed continuation; schedule settings remain unchanged.

Outcome and scope

Accept a nonempty literal string list in jobs.<id>.runs-on through the real wsr inspect <file> command. Preserve the existing scalar-string contract and its output. This is the next small extension of literal runner requirements in the inspection milestone, following the frontend modeling guidance. It makes label-based workflows inspectable without selecting a runner backend.

Example input:

jobs:
  build:
    runs-on: [self-hosted, linux, x64]
    steps:
      - id: compile
        run: echo compile

Before: exit 1, no stdout, a string-shape diagnostic at the runner field. After: exit 0, empty stderr, including:

Job: build
  Runner labels: ["self-hosted", "linux", "x64"]
  Step 1 [compile]: run

Keep the existing workflow header and execution-compatibility disclaimer. Use quoted/escaped list elements so punctuation and control characters cannot obscure label boundaries. A single-element list remains a list in the output; a scalar retains Runner: ubuntu-latest. Preserve list order, exact spelling and repetitions; do not normalize, deduplicate, infer an OS, resolve labels, discover machines, or claim that any runner satisfies the requirements.

Runner-group objects, matrices, expressions, reusable-workflow jobs and additional trigger shapes remain explicitly unsupported. In particular, an expression in a scalar or any list member must diagnose an unevaluated dynamic value, never become a literal resolved label. Do not evaluate expressions or add a runtime TypeScript sidecar.

Minimal model and implementation boundary

Acceptance and red-green sequence

  1. Run cargo test -p wsr --test cli --test inspection --locked on merged main. Baseline should retain the existing 5 CLI and 17 inspection tests; inspect live counts.
  2. Add a real CLI acceptance fixture with scalar and list jobs, identified/unidentified command and action steps, and a static needs edge. Observe the intended list-shape rejection. Implement strict decoding, summary preservation and deterministic output together.
  3. Add invalid-shape/blank-list/member/expression cases, observing meaningful behavioral red where the current implementation accepts an invalid value. Already-rejected cases are regression evidence, not invented red evidence. Implement indexed semantic diagnostics.
  4. Verify order, punctuation/escaping, exact case, repeated labels, a single-element list, byte-for-byte existing scalar output, and dependency/step-ID behavior. Include source/model assertions so preserving scalar-versus-list form is tested independently of display.
  5. Retain malicious-run/action-path and repository sentinel assertions. Inspection must read only the supplied workflow, never execute scripts, fetch/open actions, provision tools, inspect host capabilities or mutate repository state. Use temporary directories for fixtures.
  6. Review the complete diff; run focused tests, cargo xtask ci, Markdown link checks and git diff --check. Update CLI/status/testing documentation for verified behavior only.
  7. Deliver one feature PR from current origin/main, using the existing checkout. Merge only the tested head after required hosted checks and relevant review findings are resolved. Capture old/new CLI output and source-model comparison against the merged binary.

Do not broaden this slice to unblock the full example workflows: their other unsupported fields must still fail explicitly. Canonical example-*.yml acceptance remains a later milestone gate. No hooks, environment refresh, dependency churn or unrelated repairs are prerequisites.

Cost and continuity

Planning estimate for implementation through PR review: 8,000–16,000 additional tokens, with roughly 2,000–4,000 more reserved for hosted-check follow-up and merge verification. This is a rough workload estimate, not measured billing or a conversion of weekly percentage into tokens. Unexpected regressions or review findings can exceed it.

Before implementation, record live revision/branch/dirty state and any supported account-limit observation. Do not infer reset availability from old cached values. No reset is authorized by this plan, and no supported agent redemption control has been established. Keep reset management in T3; do not patch it or add repository scripts. If interrupted, checkpoint exact owned files, red/green evidence, PR head/check state and next action. Never restart completed work or spend credits automatically. Stop after this iteration; the remaining inspection milestone is not complete.